SSL-InsNet: A Sequence Structured Labeling Framework for Large-Scale Genomic Insertion Detection Leveraging Time-Distributed Dual Transformer.

Genomic structural variations (SVs) are pivotal in pathogenesis, yet existing methods often fail to maintain structural consistency for ultra-long variants spanning extensive intervals. In this paper, we propose SSL-InsNet, a novel sequence structured labeling framework that reformulates large-scale genomic insertion detection as a structured tagging task. Within a dual-transformer architecture, we develop a Multi-level Spatial Perception (MSP) module that integrates a Spatial Transformer (S-Trans) with deep convolutions via a hierarchical attention mechanism to simultaneously resolve micro-scale breakpoint motifs and regional genomic contexts. To suppress alignment artifacts, a Confidence-gated Feature Integration (CFI) module is introduced to adaptively fuse these heterogeneous spatial features based on learnable descriptors. Concurrently, a Time-distributed Transformer (T Trans) models macro-scale long-distance dependencies along the temporal axis across consecutive segments, ensuring global structural coherence. To optimize the sequential tags, we propose the Neighborhood-Weighted Focal Loss (NWFL) to explicitly enforce biological spatial continuity and penalize frag mented predictions. Comprehensive benchmarking on multiple long-read sequencing datasets demonstrates that SSL-InsNet consistently achieves state-of-the-art performance. Notably, for large-scale insertions, our framework effectively mitigates signal fragmentation and yields a superior balance between precision and recall, outperforming competitive deep learning baselines by a significant margin.